Skip to content

16-bit Support and Dynamic Loss Scaling #360

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
jasonkena wants to merge 10 commits into
base: master
Choose a base branch
from
Open

16-bit Support and Dynamic Loss Scaling #360

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
43204d1
Formatted with Black
jasonkena Feb 23, 2020
169ac9c
APEX Amp Support
jasonkena Feb 24, 2020
2a778e9
Backup before merge with remote
jasonkena Feb 27, 2020
834b7b0
Second backup
jasonkena Feb 27, 2020
f6d5d49
Clean
jasonkena Feb 27, 2020
2584d23
TACO Dataset
jasonkena Feb 27, 2020
5db6856
AMP Working but multibox_loss broken
jasonkena Feb 27, 2020
5e1be29
Everything AMP Working
jasonkena Feb 27, 2020
9dd4233
Updated README
jasonkena Feb 27, 2020
e1a9494
Fixed README
jasonkena Feb 27, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 10 additions & 5 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,10 +2,10 @@
```
██╗ ██╗ ██████╗ ██╗ █████╗ ██████╗████████╗
╚██╗ ██╔╝██╔═══██╗██║ ██╔══██╗██╔════╝╚══██╔══╝
╚████╔╝ ██║ ██║██║ ███████║██║ ██║
╚██╔╝ ██║ ██║██║ ██╔══██║██║ ██║
██║ ╚██████╔╝███████╗██║ ██║╚██████╗ ██║
╚═╝ ╚═════╝ ╚══════╝╚═╝ ╚═╝ ╚═════╝ ╚═╝
╚████╔╝ ██║ ██║██║ ███████║██║ ██║
╚██╔╝ ██║ ██║██║ ██╔══██║██║ ██║
██║ ╚██████╔╝███████╗██║ ██║╚██████╗ ██║
╚═╝ ╚═════╝ ╚══════╝╚═╝ ╚═╝ ╚═════╝ ╚═╝
```

A simple, fully convolutional model for real-time instance segmentation. This is the code for our papers:
Expand Down Expand Up @@ -44,7 +44,7 @@ Some examples from our YOLACT base model (33.5 fps on a Titan Xp and 29.8 mAP on
```Shell
# Cython needs to be installed before pycocotools
pip install cython
pip install opencv-python pillow pycocotools matplotlib
pip install opencv-python pillow pycocotools matplotlib
```
- If you'd like to train YOLACT, download the COCO dataset and the 2014/2017 annotations. Note that this script will take a while and dump 21gb of files into `./data/coco`.
```Shell
Expand Down Expand Up @@ -169,6 +169,11 @@ YOLACT now supports multiple GPUs seamlessly during training:
- If you have memory to spare you can increase the batch size further, but keep it a multiple of the number of GPUs you're using.
- If you want to allocate the images per GPU specific for different GPUs, you can use `--batch_alloc=[alloc]` where [alloc] is a comma seprated list containing the number of images on each GPU. This must sum to `batch_size`.

## 16-bit Precision Support and Dynamic Loss-scaling
YOLACT now supports the use of [NVidia's Apex AMP](https://github.com/NVIDIA/apex), enabling computation in FP16, while maintaining the weights in FP32. The use of dynamic loss scaling also prevents the `Moving average ignored a value of inf/nan` error. The only drawback is that [Apex does not support `torch.jit`](https://github.com/NVIDIA/apex/issues/303).

To enable Apex AMP support, set `use_amp` in `data/config.py` to `True`.

## Logging
YOLACT now logs training and validation information by default. You can disable this with `--no_log`. A guide on how to visualize these logs is coming soon, but now you can look at `LogVizualizer` in `utils/logger.py` for help.

Expand Down
Loading